claude-mem hits 65.8K stars as a persistent memory plugin for Claude Code

One of the most consistent complaints I hear from developers using AI coding tools is that they forget everything between sessions. Every new conversation starts cold. You re-explain the project structure, the recent decisions, the ongoing bugs, and then do it all again tomorrow.

claude-mem is the most practical solution to that problem I've come across. The plugin automatically captures what Claude does during coding sessions, compresses observations using AI, and injects relevant context back into future sessions, giving Claude Code a form of long-term memory. It just crossed 65.8K GitHub stars, which tells me this pain point is much more widespread than vendors would like to acknowledge.

What Happened

Developer Alex Newman (@thedotmack) released claude-mem as a Claude Code plugin that records tool usage observations, generates semantic summaries, and stores them in a local SQLite database with ChromaDB vector search. The project reached v12.3.8 as of April 20, 2026, with 1,792 commits, 106 contributors, and 244 releases.

What I'd flag here is the breadth of support. It works with Claude Code, Gemini CLI, and OpenCode, and installs with a single command: npx claude-mem install. The plugin also runs a local worker service on port 37777, with a web-based UI, 10 search endpoints, and an MCP server with dedicated search tools.

What I find notable is that much of the plugin's own development was done with Claude. Commit messages throughout the repo show co-authorship with Claude Opus 4.5, 4.6, and 4.7, making it a visible example of AI-assisted open source development at scale.

Key Features

• Automatic session capture. Five lifecycle hooks (SessionStart, UserPromptSubmit, PostToolUse, Stop, SessionEnd) record everything Claude does without manual intervention. You don't have to remember to log anything.
• AI-compressed observations. Raw session data is compressed via Claude's agent-sdk into structured observations containing facts, concepts, and file references, which are stored in SQLite. The compression is what keeps token costs manageable.
• Hybrid search with ChromaDB. Combines FTS5 full-text keyword search with Chroma vector embeddings for semantic retrieval across project history. Local embeddings mean no external API calls needed.
• Token-efficient 3-layer MCP workflow. Search returns compact indexes, then fetches full details only for filtered IDs, roughly 10x token savings over naive retrieval. That efficiency matters on long-running projects.
• Privacy controls. <private> tags exclude sensitive content from storage. The worker binds to localhost only, with rate-limiting and path-traversal protection.
• Multi-IDE support. Works with Claude Code, Gemini CLI, and OpenCode, with Docker containers for isolated testing and SWE-bench evaluation scaffolding.

Why It Matters

AI coding assistants lose all project context when a session ends. Developers end up re-explaining architecture decisions, repeating bug context, and watching the AI re-discover patterns it already found yesterday. claude-mem addresses this directly by creating a persistent knowledge layer that survives session boundaries.

For teams, the practical result is real. Claude Code can recall prior debugging sessions, remember which files were modified in a feature branch, and pick up refactoring work without a cold start. The progressive disclosure approach, injecting only relevant compressed context rather than full transcripts, keeps token costs manageable while maintaining continuity.

The broader pattern I'm seeing is that developers are building infrastructure around AI assistants rather than waiting for vendors to ship it. claude-mem is one of the clearer expressions of that shift, and 65.8K stars suggest it's resonating.

Example Use Case

A developer working on a Node.js API spends a session with Claude Code debugging a race condition in their authentication middleware. Claude reads several files, identifies the issue in session handling, and applies a fix. The session ends.

Two days later, the developer starts a new Claude Code session in the same project. claude-mem's SessionStart hook fires, queries its SQLite and ChromaDB store for observations tagged to this project, and injects a compressed summary into Claude's context. Claude now knows about the race condition fix, which files were involved, and the architectural decisions made, without the developer typing a word of explanation.

This is the workflow I'd demo to anyone skeptical about whether persistent memory actually changes how you work. It does.

Competitive Context

Claude Code ships with CLAUDE.md files for persisting project notes, but they're static markdown files that developers must write and maintain manually. claude-mem automates the capture, compress, and retrieve loop entirely, removing the manual overhead.

Claude's built-in conversation memory operates at the account level and stores general preferences. claude-mem operates at the project level, storing granular technical observations: which files were read, what bugs were found, what patterns were applied. The two serve different purposes. claude-mem fills the gap between Claude Code's session-scoped context window and the kind of persistent, searchable project memory that multi-session coding work requires.

For Gemini CLI and OpenCode users, claude-mem provides the same persistent memory layer, making it one of the few tools that spans multiple AI coding environments. That cross-ecosystem support is underrated for teams that don't want to commit entirely to one vendor.

My Take

claude-mem adds persistent, searchable project memory to Claude Code sessions with zero manual effort after install. If you use Claude Code across multiple sessions on the same codebase, especially for debugging, refactoring, or long-running feature work, it directly addresses the context loss problem.

Install with npx claude-mem install and restart Claude Code. At 65.8K stars, 5.6K forks, and 106 contributors, the community weight behind it is real. I'd rather adopt a tool that's already been stress-tested across this many codebases than wait for the problem to get solved at the vendor level.